Cognitive Technologies in 5G Slicing Management and Control Eugen Borcoci University Politehnica Bucharest Electronics, Telecommunications and Information Technology Faculty ( ETTI) [email protected]Slide 1 InfoWare 2019 Conference, June 30 - July 04, Rome
113
Embed
Cognitive Technologies in 5G Slicing Management and Control · 2019-09-13 · Cognitive Technologies in 5G Slicing Management and Control Eugen Borcoci University Politehnica Bucharest
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Cognitive Technologies in 5G Slicing
Management and Control
Eugen BorcociUniversity Politehnica Bucharest
Electronics, Telecommunications and Information Technology Faculty
provider/operator context many open research issues and challenges
5G slicing management and control (M&C) aspects
Service/data model & mapping on slices
Customized slice design and preparation, stitching / composition in a
single domain and cross-domain
Cognitive Technologies in 5G Slicing Management and Control
Slide 4
InfoWare 2019 Conference, June 30 - July 04, Rome
single domain and cross-domain
Network slice life cycle management, monitoring and updating
M&C system: should react in real-time, based on complex
decision making techniques, that analyse historical, temporal and frequency network data
• Cognitive network management technologies added to M&C,
allows : self-aware, self-configuring, self-optimization, self-healing and self-protecting characteristics
• Generally, self – organizing – networks (SON) capabilities can be achieved
CONTENTS
1. Introduction
2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Slide 5Slide 5
InfoWare 2019 Conference, June 30 - July 04, Rome
CONTENTS
1. Introduction
2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Slide 6Slide 6
InfoWare 2019 Conference, June 30 - July 04, Rome
1.1 5G general aspects
Three views/sets-of-requirements for 5G user-centric, service-provider-centric, network-operator-centric
5G Key technological characteristics Integrates different and heterogeneous access technologies, cellular,
Radio Access Technologies (RAT), sattellites, .. Ultra-dense networks with numerous small cells
Driven by SW unified OS in a number of PoPs, especially placed at the network
1. Introduction
Slide 7
InfoWare 2019 Conference, June 30 - July 04, Rome
unified OS in a number of PoPs, especially placed at the network edge
Complementary technologies for 5G slicing Software Defined Networking (SDN) Network Functions Virtualization (NFV) Cloud/Mobile Edge Computing (MEC) /Fog Computing (FC)
Recent trends: Optimized and advanced M&C cognitive features, autonomic management advanced automation of operation through proper algorithms Data Analytics and Big Data techniques -> monitor the users’ QoE
Slide 7
1.1 5G general aspects
Network softwarization
programmability of network devices
network functions (NF)- virtual or physical
network slices – logical, on demand, customized networks
softwarization capabilities in all network segments and network components
1. Introduction
Slide 8
InfoWare 2019 Conference, June 30 - July 04, Rome
Separation of concerns between
control/ management/ services
logical / physical resources functions (in terms of connectivity, computing and
storage)
On demand composition of NFs and network capabilities
See: A.Galis, 5G Architecture Viewpoints H2020 5G PPP Infrastructure Association July 2016, August
2017, https://5g-ppp.eu/white-papers/
1.2 5G Key Specific Requirements
Summary of 5G figures - strong goals 1,000 X in mobile data volume per geographical area reaching a target ≥ 10 Tb/s/km2
1,000 X in number of connected devices reaching a density ≥ 1M terminals/km2
100 X in user data rate reaching a peak terminal data rate ≥ 10Gb/s
1. Introduction
Slide 9
InfoWare 2019 Conference, June 30 - July 04, Rome
100 X in user data rate reaching a peak terminal data rate ≥ 10Gb/s
1/10 X in energy consumption compared to 2010
1/5 X in E2E latency reaching 5 ms for e.g. tactile Internet and radio link latency reaching a target ≤ 1 ms, e.g. for Vehicle to Vehicle (V2V) communication
1/5 X in network management OPEX
1/1,000 X in service deployment time, reaching a complete deploymentin ≤ 90 minutes
Slide 9
1.3 5G Generic Architecture multi-tier architecture: small-cells, mobile small-cells, D2D- and Cognitive Radio
Network (CRN)
1. Introduction
DR-OC - Device relaying with operator controlled link establishment
DC-OC - Direct D2D communication with operator controlled link establishment
DR-DC - Device relaying with device controlled link establishment
DC-DC - Direct D2D communication with device controlled link establishment
Slide 10
InfoWare 2019 Conference, June 30 - July 04, Rome
Slide 10
Source: Panwar N., Sharma S., Singh A. K. ,A Survey on 5G: The Next Generation of Mobile Communication’.
Accepted in Elsevier Physical Communication, 4 Nov 2015, http://arxiv.org/pdf/1511.01643v1.pdf
1.4 4G versus 5G concepts
1. Introduction
Slide 11
InfoWare 2019 Conference, June 30 - July 04, Rome
MBB - Mobile Broadband SMS - Short Messages service LTE - Long Term Evolution (4G) EPC- Evolved Packet CoreV2X - vehicle to X ; CNF- Core Network Functions RNF- RAN network Functions
1.5 Network slicing concepts
E2E concept: covering all network segments: radio, access/edge, wire, core,
transport and edge networks.
concurrent deployment of multiple E2E logical, self-contained and
independent shared or partitioned networks on a common infrastructure
Slices created by provisioning/ on_demand, isolated (w.r.t. performance,
security), each one with its independent M&C
1. Introduction
Slide 12
InfoWare 2019 Conference, June 30 - July 04, Rome
security), each one with its independent M&C
composition of adequately configured NFs, network apps., and the
A.Galis and K.Makhijani, Network Slicing Landscape: A holistic architectural approach, orchestration and management
with applicbility in mobile and fixed networks and clouds, v1.0, Network Slicing Tutorial – IEEE NetSoft 2018 – Montreal 29th
June2018.
1.8 Summary of Network Slices key requirements (cont’d)
Business and network operator/ SP
NSL should support
open possibility of new business models industrial companies can use NSs as a part of their own services
reduced operations expenditures (OPEX)
programmability allows to enrich the offered services
1. Introduction
Slide 21
InfoWare 2019 Conference, June 30 - July 04, Rome
programmability allows to enrich the offered services
OTT providers and other market players can use NSLs without changing the PHY infrastructure
to simplify the provisioning of services, manageability, integration and
operation
to create a layer of abstraction by the creation of L/P isolated groups of
network resources and VNFs
isolation, orchestration and separation of logical network behaviors from the
underlying PHY network resources.
See:. L. Geng , et.a;., IETF- “Network Slicing Architecture draft-geng-netslices-architecture-02”, 2017A.Galis and K.Makhijani, Network Slicing Landscape: A holistic architectural approach, orchestration and management with applicbility in mobile and fixed networks and clouds, v1.0, Network Slicing Tutorial – IEEE NetSoft 2018 – Montreal 29th June2018.
CONTENTS
1. Introduction
2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Solutions defined by 3GPP-for Core Network (CN) slicing (cont’d)
Group B - some NFs are common between the NSLs, while other
functions reside in individual NSLs
Group C - the Control Plane is common between the slices, while the
User plane(s) (UPl/DPl) are handled as different NSLs
2. 5G slicing relevant architectures
Slide 24
InfoWare 2019 Conference, June 30 - July 04, RomeSee: 3GPP, TR 23.799 V14.0.0 (2016-12), Study on Architecture for Next Generation System (Release 14)
Note: This approach does not detail the slicing at AN premises
2.2 Slicing variants - examples
2. 5G slicing relevant architectures
CN
Communication
Services
CN CN CN
NSI A NSI B NSI C
Service 1 Service 2 Service 3
Network slice subnet instance C
NF5 NF6
Network slice instance X
End to End services provided by NSLI(s) One NSSI can contribute to several NSLIs
E.g., NSLI X and Y composed by NSSI A, B and C
Slide 25
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: 3GPP TR28.801 V15.1.0 (2018-01), Study on management and orchestration of network slicing for next
generation network, (Release 15)
AN
CN
AN NSSI 1
CN NSSI 2
CN NSSI 1
AN NSSI 2
CN NSSI 3
Network slicesubnet instance A
NF1 NF2 NF4
NF3
Network slice subnet instance C
Network slicesubnet instance B
NF7
NF9
NF8
Network slice instance Y
2.3 ETSI and 3GPP functional architectures for slicing support Network slice management (NSLM) in NFV framework
2. 5G slicing relevant architectures
New slicing mgmt.
components added to NFV
framework
Slide 26
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: ETSI GR NFV-EVE 012 V3.1.1 (2017-12), Release 3; NFV Evolution and Ecosystem; Report on
Network Slicing Support with ETSI NFV Architecture Framework
Relating the information models
2.3 ETSI and 3GPP functional architectures for slicing support Network slice management (NSM) in NFV framework (cont’d)
Three layered functions related to NSL mgmt.
Communication Service Management Function (CSMF): translates the
comm. service requirements to NSL requirements; I/F with (NSMF)
Network Slice Management Function (NSLMF) – management
(including lifecycle) of NSLIs
It derives NSL subnet requirements from the NSL related requirements
I/F with NSSMF and the CSMF
Network Slice Subnet Management Function (NSSMF) – mgmt
2. 5G slicing relevant architectures
Slide 27
InfoWare 2019 Conference, June 30 - July 04, Rome
Network Slice Subnet Management Function (NSSMF) – mgmt
(including lifecycle) of NSSIs.
I/F with the NSMF
The Os-Ma-NFVO Reference Point (RP) is the I/F with NFV-MANO
The NSMF and/or NSSMF have to
determine the type of NSL or set of NSLs, VNF and PNF that can
support the resource requirements for a NSLI or NSSI
analyze if existing instances can be re-used, else need to create new
instances of these NSLs, VNFs and the connectivity to the PNFs
See ETSI GR NFV-EVE 012 V3.1.1 (2017-12), Release 3; NFV Evolution and Ecosystem; Report on Network
Slicing Support with ETSI NFV Architecture Framework
Network Function Virtualisation – ETSI- summary
High level view of NFV framework ( recall)
Working domains composed of VNF -SW implementation of a NF which is running over the NFVI NFV Infrastructure (NFVI), including the diversity of physical resources
and virtualisation tools NFVI supports the execution of the VNFs
The Virtualisation Layer (VL) abstracts the HW resources and
NFV summary 1/2 (complementary slide)
Slide 28
InfoWare 2019 Conference, June 30 - July 04, Rome
The Virtualisation Layer (VL) abstracts the HW resources and decouples the VNF software from the underlying hardware, thus ensuring a HW-independent lifecycle for the VNFs
NFV Management and Orchestration (NFV-MANO) orchestration and lifecycle management (LCM) of physical and/or
SW resources that support the infrastructure virtualisation, and the VNFs lifecycle management
NFV MANO focuses on management of all virtualisation-specific tasks
See: ETSI GS NFV 002 v1.2.1 2014-12, NFV Architectural Framework
Network Function Virtualisation – ETSI- summary (cont’d)
NFV Management and Orchestration Architectural Framework (NFV-MANO Architectural Framework): collection of all FBs (those in NFV-MANO and others interworking with
NFV-MANO), data repositories used by these FBs , and Reference Points
(RPs) and interfaces for the purpose of managing and orchestrating NFV
NFV summary 2/2 (complementary slide)
Slide 29
InfoWare 2019 Conference, June 30 - July 04, Rome
Network Functions Virtualisation Orchestrator (NFVO): FB that manages the Network Service (NSrv) lifecycle and coordinates
the management actions for NSrv lifecycle
VNF lifecycle (supported by the VNFM)
NFVI resources (supported by the VIM)
to ensure an optimized allocation of the necessary resources and
connectivity
See: ETSI GS NFV 002 v1.2.1 2014-12, NFV Architectural Framework
2.3 5G Layered Architecture - 5GPPP vision
2. 5G slicing relevant architectures
Slide 30
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: 5GPPP Architecture Working Group , View on 5G Architecture, Version 2.0, December 2017
Note: The name “layer” here does not correspond to the classic one; the correct semantic would be rather - “plane”
2.3 5G Layered Architecture - 5GPPP vision (see the previous slide) Architecture based on ETSI-NFV and SDN
Service layer
Apps. and services operated by the tenant (includes the E2E orch. system)
Business Support Systems (BSSs); Business-level Policy and Decision functions
Operation Support System
Management and Orchestration layer
Service Management (i.e., services offered by the slices)
Software-Defined Mobile Network Orchestrator (SDM-O)
managed by Wide Area Infrastructure Managers (WIMs)
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d)
The NSL provider can simultaneously operate multiple NSLIs rents the infrastructure resources owned by the InPs
NSLs are mutually isolated: w.r.t performance, resiliency, security, privacy and management they run concurrently on top of a shared NFVI without (directly or
indirectly) affecting each other
the infrastructure (and NFVI) is owned and managed by different (and
Slide 34
InfoWare 2019 Conference, June 30 - July 04, Rome
the infrastructure (and NFVI) is owned and managed by different (and potentially non-trusted) administrative domains (InP1, InP2, ..)
Details on isolation: e.g., ETSI GS NFV-EVE 005, R NFV-IFA 022 and GR NFV-IFA 028
NSL Orchestrator (NSLO)- highest layer of the architecture key role in the creation phase and also in the run-time phase
3. Management, orchestration and control
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d) NSLO role at creation phase
It receives the order to deploy a NSLI for a tenant (or, the Slice Provider decides to construct a slice)
It has information (including on multi-domain) as to check the feasibility of the order interacts with RO and accesses the VNF and NS Catalogues the catalogues contain VNF and NS descriptors, exposing the
capabilities of all the VNFs and NSs that an NSL provider can select for the NSLs
if feasible order, then NSLO triggers the instantiation of the NSL
Slide 35
InfoWare 2019 Conference, June 30 - July 04, Rome
if feasible order, then NSLO triggers the instantiation of the NSL
NSLO role at run-time phase performs policy-based inter-slice operations
e.g., it analyzes the performance and fault management data received from the operative NSLIs instances, to manage their SLAs
If SLA violations, the NSLO decides to modify/correct some NSLIs
Resource Orchestration (RO) uses the resources (supplied by the VIMs/WIMs) and dispatches them to
the NSL instances in an optimal way To do this, it needs to know the resource availability in each domain
(this supposes a set of inter-domain interactions)
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d)
A NSL instance (NSLI) may be composed of one or more Network Service (NS) instances
instance of a simple NS instance of a composite NS concatenation of simple and/or composite NS instances
can span several Infrastructure Providers (InP) and/or admin. domains has its own MPl and CPl planes and this provides isolation across NSIs
NSL Manager Network Service Orchestrator (NSO)
Slide 36
InfoWare 2019 Conference, June 30 - July 04, Rome
Network Service Orchestrator (NSO) Tenant SDN Controller (TSC) VNF Manager (VNFM)
The VNFM(s) and the NSO perform the required life cycle operations (e.g., instantiation, scaling, termination, etc.) over the instances of the VNFs and NS(s), respectively
NSL Manager – key element for a NSL tenant coordinates the O&M data, from both NSO and TSC performs the FCAPS – set of functions within the NSLI provides visibility and mgmt. capability exposure to external blocks establishes the limits within each tenant may operate and consume
its NSLI instance
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d) The VNFM(s) and the NSO perform
life cycle operations (e.g., instantiation, scaling, termination, etc.) over the instances of the VNFs and NS(s), respectively
these operations involve modifying the amount of resources to be allocated for those instances an interaction with RO is needed
Tenant SDN Controller (TSC) dynamically configures and chains VNFs to realize network services
(NS) in the tenant domain (TSC can be deployed as a VNF itself) it chains the VNF instances for NS construction, leveraging the
Slide 37
InfoWare 2019 Conference, June 30 - July 04, Rome
it chains the VNF instances for NS construction, leveraging the forwarding DPl capabilities
It configures the VNF instances at application level but not their underlying NFVI resource (it creates an overlay of VNFs)
It plays the role of an Element Manager (EM) ( see ETSI NFV framework)
It offers a set of dedicated northbound I/Fs that allows slice's clients (and thus tenant's clients) to interact with the slice
NSrvs and VNF operations are highly correlated after a NS has been instantiated, the OSS (an SDN app. from the TSC perspective), will instruct TSC to perform the VNF configuration and chaining tasks
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d)
NFVI level NFV and SDN solutions are applied M&C: VIM, WIM, SDN Infrastructure controllers
Each NFVI-PoP has a single VIM instance to configure and manage the virtualization containers and their underlying HW
Their connectivity is locally enforced by the infrastructure SDN controller (IC)
Slide 38
InfoWare 2019 Conference, June 30 - July 04, Rome
controller (IC)
To connect NFVI-PoPs, each WAN domain relies on a WAN Infrastructure Manager (WIM) instance, (similar to the model in ETSI GR NFV-IFA 022)
The scenario presented above is well-aligned with the NFVI as a Service (NFVIaaS) (ETSI GR NFV-IFA 028)
Each tenant uses the NFVI to get the performance needs of the slices in its domain each InP plays the NFVIaaS provider role each tenant acts as an NFVIaaS consumer
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d)
A tenant, with its own set of NSLs is isolated from others both VIMs and WIMs support multi-tenancy by offering separate NFVI resources to subscribed tenants through
dedicated I/Fs
VIMs has a resource pooling mechanisms to provide subscribed tenants with isolated resource environments endowed with high availability fault resilience features to support the tenant VNFs deployment
Slide 39
InfoWare 2019 Conference, June 30 - July 04, Rome
fault resilience features to support the tenant VNFs deployment
WIMs have similar mechanisms (e.g., those of the ONF TR 527) to simultaneously manage a number of virtual topologies in the WAN
with different levels of abstraction
See: ETSI GR NFV-EVE 012 V3.1.1 (2017-12), Release 3; NFV Evolution and Ecosystem; Report
on Network Slicing Support with ETSI NFV Architecture Framework
2. 5G slicing relevant architectures
2.4 Multi-tenant – multi-domain architectures – ETSI (cont’d)
The infrastructure SDN controller (IC) M&C for the NFVI resources (placed in a NFVI-PoP or a WAN)
set up the connectivity to support the communication between the tenant VNFs and/or PNFs in the infrastructure domain
performs M&C of the connectivity among the virtualization containers that host the tenant VNFs' software applications
Slide 40
InfoWare 2019 Conference, June 30 - July 04, Rome
the networking resources, supporting VM (and hence VNF) connectivity at the infrastructure level, are managed by the ICs following the managers VIM and the WIM commands
VIMs and WIMs act as SDN applications, delegating to ICs the M&C tasks related to networking resources
Implementation variant: to integrate ICs into their corresponding VIMs
See: ETSI GR NFV-EVE 012 V3.1.1 (2017-12), Release 3; NFV Evolution and Ecosystem; Report
on Network Slicing Support with ETSI NFV Architecture Framework
CONTENTS
1. 5G Network slicing – concepts, use cases and
requirements2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Slide 41
5. Conclusions and research challenges
Slide 41
InfoWare 2019 Conference, June 30 - July 04, Rome
3.1 General requirements for 5G slicing management, orchestration and
control
To support flexible business models (various actors)
Ability to create/support-- on demand/provisioned slices
in multi-tenant, multi-domain, multi-operator, E2E environments
3.4 Generic service management and network slice control- example (cont’d)
Roles of the planes Service management plane- performs service operations
abstraction, negotiation, admission control and charging for verticals
and 3rd parties
service creation if admission control accepts the slice request
AC input parameters: {slice reqs., slice templates available}
The desired service combines
VNFs, PNFs, value added services
3. Management, orchestration and control
Slide 49
InfoWare 2019 Conference, June 30 - July 04, Rome
VNFs, PNFs, value added services
data/control plane, and security mechanisms
and exposes them to the underlying network
Network slice management and control plane provides resource abstraction to service management
handles NSL resource management & control plane operations, including
instantiation of the slice resources based on the service mapping
performance maintenance via monitoring, analysis and slice re-
configuration procedures
slice selection, attachment and support for multi-slice connectivity
3.4 Generic service management and network slice control (cont’d)
Actions (related to previous slide)
(1) Network Slice M&C provides resource abstraction to service management
(2-3) abstraction, negotiation, AC and charging for verticals and 3rd parties
(4-5) service creation after a slice request is accepted by AC
3. Management, orchestration and control
Slide 50
InfoWare 2019 Conference, June 30 - July 04, Rome
(6) Srv. Mgmt provides all information to network slice M&C plane
(7) Net M&C instantiates the slice resources based on service mapping
(8) Net M&C performs slice selection, attachment and support for multi-slice
connectivity
(9) Net M&C - performance maintenance via monitoring, analysis and slice
re-configuration procedures
CONTENTS
1. Introduction
2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Slide 51Slide 51
InfoWare 2019 Conference, June 30 - July 04, Rome
4. Cognitive technologies in 5G slicing M&C
4.1 Cognitive Management concepts
5G slicing – has complex management requirements related to multi –
tenant/domain/operator context and softwarization of network resources
Need of real-time mgmt. based on a hierarchy of complex decision making
techniques that analyze historical, temporal and frequency network data
Cognitive network management – recent trend using Artificial Intelligence
(AI) and in particular Machine Learning (ML) to develop self-x, (-x= -aware, -
Slide 52
(AI) and in particular Machine Learning (ML) to develop self-x, (-x= -aware, -
configuring, -optimization, -healing and -protecting systems)
Cognitive management– extension of Autonomic Management (AM) (coined by
IBM ~ 2001)
AM + Machine learning = Cognitive Management (CogM)
Challenge: to deploy the CogM and its orchestration across multiple
heterogeneous networks: Radio & Other Access Networks, Core & Aggregation,
Edge Networks, Edge and Computing Clouds and Satellite Networks
Example 1 of ML: Supervised learning k-Nearest Neighbors (k-NN) – simple method
Assumption: the features used to describe the domain points are relevant to their labeling in a way that makes close-by (Euclidian distance) points likely to have the same label
k-NN figures out a label on any test point without searching for a
Slide 69
k-NN figures out a label on any test point without searching for a predictor within some predefined class of functions
Idea: memorize the training set, then predict the label of any new instance/data_point on the
basis of the labels of its closest neighbors in the training set a new data_point is classified by a majority vote of its k-
nearest neighbors
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: S. Shalev-Shwartz and S.Ben-David, Understanding Machine Learning: From Theory to Algorithms
2014, Cambridge University Press
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 1 of ML: Supervised learning (cont’d) k-Nearest Neighbors (k-NN) – classification example
Let it be an instance domain, X and “points”
Define :
Let
instance domain, X,
Slide 70
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: S. Shalev-Shwartz and S.Ben-David, Understanding Machine Learning: From Theory to Algorithms
2014, Cambridge University Press
ExampleClassification of the test Star point :
K=3 Class BK= 6 Class A
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 1 of ML: Supervised learning k-Nearest Neighbors (k-NN) – (cont’d)
k-NN can be also used for regression applications
Pros: Simple to implement
Works well in practice
Slide 71
Works well in practice
Does not require to build a model, make assumptions, tune parameters
Can be extended easily with news examples
Cons: Requires large space to store the entire training dataset.
Slow! Given n examples and d features. The method takes O(n x d) to run
Dimensionality problem
InfoWare 2019 Conference, June 30 - July 04, Rome
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 2 of ML: (Artificial) Neural Networks (ANN) – summary ANN - computational model inspired by the biological neural networks
in the human brain
The basic unit of computation: neuron (node) The node applies an activation (non-linear) function f to the weighted sum
of its inputs (including a bias)
The bias provides every node with a trainable constant value (in
addition to the normal inputs) Examples of activation functions
Slide 72
addition to the normal inputs)
InfoWare 2019 Conference, June 30 - July 04, Rome
Examples of activation functions
Sigmoid: σ(x) = 1 / (1 + exp(−x))
tanh: tanh(x) = 2σ(2x) − 1
ReLU: Rectified Linear Unit. f(x) = max(0, x)
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 2 of ML: Artificial Neural Networks (ANN) – summary (cont’d) Feed-forward Neural Network
Simplest type of ANN containing multiple neurons (nodes) arranged in layers
Multi Layer Perceptron (MLP) –has one or more hidden layers
Slide 73
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: Ujjwalkarn, A Quick Introduction to Neural Networks, https://ujjwalkarn.me/2016/08/09/quick-intro-neural-
networks/, 2016
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 2 of ML: Artificial Neural Networks (ANN) – summary (cont’d) Feed-forward Neural Network
Given a set of features X = (x1, x2, …) and a target y, a MLP can learn the
relationship between the features and the target, e.g., for classification or
regression
Training the MLP: The Back-Propagation (BP) Algorithm BP of errors - is one way to train an ANN
BP is a supervised training scheme
� � � �
� � � �
Slide 74
BP is a supervised training scheme
Learning a function �(�) to map given inputs � to desired outputs �
Training with 'labeled' data: each example input �(�) has a label �(�)
('correct' output)
The error � between ��(�(�)) and �(�) is used to adapt �
and compute ��+1(�)
Method: gradient based adjustment of perceptron weights to correct
errors
After training, one can use ���� for unlabeled data
InfoWare 2019 Conference, June 30 - July 04, Rome
4. Cognitive technologies in 5G slicing M&C
4.3 Machine Learning (ML) – summary (cont’d)
Example 2 of ML: Artificial Neural Networks (ANN) – summary (cont’d)
Convolutional Neural Networks (Deep NN)
Goal: Increasing the NN running speed
Layers Convolutional layers (CL)
Every neuron has just a very limited number of inputs to the vicinity
of a corresponding neuron in the previous layer
All neurons in a layer use the same set of weights
Slide 75
All neurons in a layer use the same set of weights
Pooling layers (PL)
Neighboring neurons are merged (max, sum, etc.)
The fully connected layer (MLP) at the end connects all split components of
layers
Learning is performed by using back-propagation
Advantage: large networks can be composed by using these building blocks
InfoWare 2019 Conference, June 30 - July 04, Rome
Source: J,Quittek, Artificial Intelligence in Network Operations and Management,
J,Quittek, Artificial Intelligence in Network Operations and Management, https://networking.ifip.org/2018/images/2018-
IFIP-Networking/Keynote-III-J-Quittek-Slides.pdf
InfoWare 2019 Conference, June 30 - July 04, Rome
CONTENTS
1. Introduction
2. 5G slicing relevant architectures3. Management, orchestration and control4. Cognitive technologies in 5G slicing M&C5. Conclusions and research challenges
Slide 98Slide 98
InfoWare 2019 Conference, June 30 - July 04, Rome
Challenges in Using Machine Learning Representative Datasets Speed vs. accuracy Ground truth (refers to the accuracy of the training set's
classification for SML techniques) Incremental Learning Security of Machine Learning
5. Conclusions and research challenges
Slide 99
InfoWare 2019 Conference, June 30 - July 04, Rome
Security of Machine Learning
Challenges in Autonomic Network Management in cognitive context Orchestration of Cognitive Management Functions Cooperation between Cognitive Mgmt and SDN, NFV environment Selection of the most convenient ML techniques for 5G M&C
Thank you ! Questions?
Cognitive Technologies in 5G Slicing Management and Control
Slide 100
InfoWare 2019 Conference, June 30 - July 04, Rome
References1. Panwar N., Sharma S., Singh A. K. ‘A Survey on 5G: The Next Generation of Mobile Communication’ Elsevier
Physical Communication, 4 Nov 2015, http://arxiv.org/pdf/1511.01643v1.pdf
2. 5GPPP Architecture Working Group , View on 5G Architecture, Version 2.0, December 2017
3. 3GPP TS 23.501 V15.2.0 (2018-06), System Architecture for the 5G System; Stage 2, (Release 15)
4. 3GPP TS 28.530 V1.2.1 (2018-07), Management and orchestration of 5G networks; Concepts, use cases and
requirements
5. 3GPP, TR 23.799 V14.0.0 (2016-12), Study on Architecture for Next Generation System (Release 14)
6. 3GPP TR 28.801 V15.1.0 (2018-01), Study on management and orchestration of network slicing for next generation
network, (Release 15)
7. A.Galis, 5G Architecture Viewpoints H2020 5G PPP Infrastructure Association July 2016, August 2017, https://5g-
ppp.eu/white-papers/
8. A. Galis, Towards Slice Networking, presentation at IETF98 -, March 2017; https://github.com/netslices/IETF-
NetSlices
Cognitive Technologies in 5G Slicing Management and Control